06 — Zoho Alignment (Turning Your Day Job Into Your Portfolio)¶

Fifteen hours a week from personal time is what you can afford. Twenty-five hours a week is what phases 4–7 need at peak. The only way that math works is if some of the roadmap runs INSIDE Zoho, on Zoho’s clock, delivering Zoho’s business goals. This file is how you do that without gaslighting yourself or your employer.

Companion doc: ../08_production/13_zoho_leverage_plan.md has the tactical four-play plan. This file frames it as time-alignment strategy over 13 months.

The Fundamental Bet¶

Zoho is unusually well-positioned to want an inference engineer. Reasons:

On-prem-first culture. CRM-adjacent, regulated industries, data-sovereignty-sensitive customers. On-prem LLM inference is a product requirement, not a preference.
Agentic products in flight. You’re already writing agent harnesses. Someone has to serve those models efficiently. That someone should be you.
Cost sensitivity. Frontier API calls at scale are murder on margins. Fine-tuned 7–14B models running on owned hardware are the escape hatch.
India + Chennai R&D presence. Zoho invests in deep R&D, not just wrapper products. This kind of infrastructure work has a home there.

The bet: your roadmap and Zoho’s serving needs are the same thing. You don’t have to hide the roadmap. You have to align it.

The Quarterly Alignment Plan¶

Q1 (M1–M3, Jul–Sep 2026): Learn Quietly¶

Your posture: Curious student on the side. Do not announce anything.

Why quietly: You have not yet earned the right to propose infrastructure directions. R1 hasn’t shipped. Your credibility on this stack is still theoretical. Talking early looks like ambition without substance.

What to do:

Ship R1 (CPU matmul writeup) publicly. If a Zoho colleague notices, that’s fine. If not, also fine.
Notice which of your current agent harnesses have long system prompts, tool-call chains, high per-request cost. Take mental notes. Do not act.
Casually read the vLLM and SGLang blogs during coffee. Absorb the vocabulary.

What NOT to do:

Do not tell your manager “I want to move to inference infra.”
Do not volunteer to “own the serving layer” before you can implement one.
Do not badmouth existing serving choices.

Q2 (M4–M6, Oct–Dec 2026): Begin Measuring¶

Your posture: Curious engineer with data. R1 and R2 are shipped or shipping.

What to do:

Play 1 seed: With permission and sanitization, start measuring prefix cache hit rates on your real agent traffic. Every multi-turn agentic loop re-sends the system prompt + history. If you’re not caching, you’re paying 40–60% cost overhead.
Q1 review conversation with your manager: Gently introduce the thesis. Frame:

“I’ve been going deep on inference infrastructure on personal time. Our on-prem story is going to need serving expertise as we scale agent products. I’d like to bring some of this back to work over the next quarter — starting with measuring where we’re leaving performance on the table.”
If the manager bites → propose a 20% time allocation. If cautious → propose a specific low-risk artifact (a benchmark of your current setup vs vLLM continuous batching).
Publish R1 and R2 externally. Do NOT hide them internally either — send to a friendly senior with a subject line like “weekend project, thought you’d find this fun.”

Warning: October Diwali + Q1 review pressures mean this is a stretch quarter. Do not force the conversation if the timing is bad.

Q3 (M7–M9, Jan–Mar 2027): Propose and Champion¶

Your posture: Practitioner. R3 and R4 are shipped. R5 in flight. This is when you become the person leadership consults on inference.

What to do:

Play 1 delivery: Prefix caching for agent loops on Zoho traffic. Measured impact: 40–60% inference cost reduction. Present it internally.
Play 2 proposal: The internal benchmark harness. Frame as: “we need to be able to compare model/engine/quant choices with numbers, not vendor claims. Here’s what I’ve built externally, here’s how I’d port it internally.” This is a political + technical play — whoever owns the harness owns the vocabulary of decisions.
Renegotiate time allocation upward. Target: 40–50% Zoho time on inference work by end of Q3. If not possible, the personal-time load remains manageable but tight.
The pitch sentence (internally): “I build agentic services on top of LLMs AND the inference infrastructure that serves them. That combination is rare, and it’s exactly what our on-prem LLM story needs.”

Q4 (M10–M13, Apr–Jul 2027): Own Institutional Infrastructure¶

Your posture: The person leadership consults on serving. R5 shipped, R6 merged, R7 in flight.

What to do:

Play 3 execution: On-prem model registry + license compliance. Legal loves you. Security loves you. Product loves you because customer deployments unblock.
Play 4 execution: Fine-tuned 7–14B champion for a specific Zoho task. 10–100× cost win vs frontier for a bounded problem. Direct measurable ROI leadership will recognize.
The retention conversation. If Zoho matches / offers a staff-engineer inference infra role → seriously consider. See ../99_pre_mortem/07_geography_visa_offer_risk.md for the honest evaluation framework.
Simultaneously, the external study loops (see ../13_discipline/09_interview_conversion.md). Not disloyalty — market discovery. Even if you stay, knowing your market rate is essential.

The Four Plays (Detailed)¶

Full tactical detail: ../08_production/13_zoho_leverage_plan.md. Compressed here:

Play	Started	Delivered	Business Impact	Career Signal
1. Prefix caching for agents	Q2	Q3	40–60% inference cost reduction	You are the RadixAttention person at Zoho
2. Internal benchmark harness	Q3	Q3–Q4	Neutral vocabulary for eng decisions	You define how the org measures
3. On-prem model registry + licensing	Q4	Q4+	Unblocks regulated-industry customers	You own institutional infrastructure
4. Fine-tuned 7–14B champion	Q4	Q4+	10–100× cost win on specific tasks	You are the fine-tuning practitioner

These are staff-engineer promotion track deliverables at any competent org. Even if the promotion doesn’t happen, the artifacts convert to study signal at external companies.

The Conversation With Your Manager (Template)¶

Have this in Q2, not Q1. After R1 is shipped and R2 is in flight. Not before.

Manager: “How’s your quarter going?”

You: “Good. I want to talk about something. Over the last few months on personal time, I’ve been going deep on inference engineering — the stack under model.generate(). CUDA kernels, engines like vLLM, quantization, distributed serving. I’ve published a few things externally [share R1, R2 links].

“The reason I’m bringing this up: our on-prem LLM story is going to need this expertise. As we scale agentic products, we’re going to hit real serving-cost and latency problems, and the fix isn’t more prompting — it’s understanding the layer below. I’d like to bring some of what I’m learning back into work.

“Specifically, I want to spend a portion of my time on three things: (1) measuring where our current agent serving leaves performance on the table, (2) building an internal benchmark harness so we can make model/engine/quant decisions with real numbers, (3) exploring on-prem-friendly quantization + fine-tuning for customer-specific tasks.

“I’ll keep delivering on my current commitments. This isn’t ‘let me switch teams.’ This is ‘let me apply what I’m learning here, because it will pay for Zoho and because it’s what I want to do long-term anyway.’”

Expect one of three responses:

Yes, full support. (Best case, ~20% probability at first ask.) Take it. Ship prefix caching within one sprint.
Yes, but show me numbers first. (Most likely, ~50%.) Prefix caching benchmark on a small internal service is your response. Get it done in one week.
Not now. (~30%.) Fine. Continue personal-time work. Reraise in Q3 with more artifacts.

If Response 3 twice → serious escape-hatch conversation. See ../13_discipline/09_interview_conversion.md.

What Aligning At Zoho Actually Buys You¶

~5–10 extra hours/week of paid roadmap work.
Real production data (prefix hit rates, agent traffic distributions, GPU utilization) that fuels R5 and R7.
Internal reputation shift — from “applied ML person” to “the inference person.”
A staff-eng promotion track if you want it.
Optionality. If you leave, you leave with real Zoho artifacts. If you stay, you stay at a higher level.

What It Does NOT Buy You¶

Immunity from the fundamentals. If you skip Phase 0–2, no amount of Zoho alignment saves you.
A short-cut on public artifacts. Zoho-internal wins don’t substitute for R5 + R6. The studies are external either way.
Permission to burn out. See health floor in 04_weekly_rhythm.md.

The Honest Failure Mode¶

The most likely way Q3–Q4 alignment doesn’t happen isn’t refusal. It’s inertia + reorg risk. Your manager says yes, then a reorg hits, then priorities shift, then the prefix-caching work becomes “deprioritized this quarter.”

If you smell that pattern by end of Q3: pre-mortem file 02 has the parry. In short: do not slow down the personal-time roadmap to wait for internal alignment. The roadmap has to survive Zoho’s fluctuations because Zoho is not the mission — the skill is.

The One Sentence¶

Q1 quiet, Q2 measure, Q3 propose, Q4 own. Your day job is either accelerant or friction; you get to choose which by how you frame the conversation.

Return to README.md · Next: 07_thirteen_month_pitch.md